Name: In Focus: Fin Guidance
Uploaded: 2025-10-02T19:01:10.433Z
Duration: 45 min 24 s
Description: In Focus: Fin Guidance

Transcript for "In Focus: Fin Guidance": Good afternoon, everyone. Thank you for coming to today's session on guidance. This is a new series of webinars that we're launching today called InFocus, which is geared towards more advanced fin users. So, this is probably why you're in attendance today. And, the whole objective is just do a deep dive on, some of our more advanced AI features. And, of course, today's subject is gonna be guidance and then how we can best, apply that to our AI agent. As always, my name is Tim, and I'm a customer success manager here at Intercom. If it's your first time joining us today, then welcome. If not, then you're probably already familiar with, the webinar layout. But for those of you, who are joining us for the first time, I'd like to walk you through the layout of of Goldcast really quickly. So on the right hand side, you've probably already seen the chat. Please feel free to pop in a greeting, say hi, let us know where you're joining in from today, Drop a gift for my colleagues backstage, and let's keep it for general comments in the chat only. For any questions that you have for us throughout the session, I've encouraged you to go to the q and a tab on the right hand side of the screen. This is where, me and my team will be answering questions both backstage and on stage, and upload any questions that you find relevant or not. The doc section is gonna have several resources that you can download throughout the session, so I highly encourage you to take a look through those and bookmark them. And then finally, today's session is being recorded and will be available on demand for anyone that's registered. So for any colleagues that weren't able to attend today, please make sure that they're registered so that they can receive a link to the recording. So without further ado, here is a quick look at what we'll be covering today. First, we're gonna do a quick pass of the guidance essentials. So how it works, how to use it, how not to use it, and then we'll go into some of the best practices that are recommended by our experts, which is gonna lead us to the demo, where some colleagues of mine will be showing you some of these practices in action within the test workspace. And then of course, we're gonna tie things off with some q and a at the end. As you've probably seen in the event description, today's session is more geared towards more advanced users. Who have already had the chance to test guidance and then set some live, which is why I don't wanna spend too much time on the basics, but rather provide some clarity on guidance overall to help you understand the logic behind it better. So how do the basics of guidance work? Guidance is applied based on context and relevancy. Meaning, the more detailed the prompt is, the better your AI agent is at understanding when it should or shouldn't apply guidance to the final response, which therefore means that the performance of guidance is based on the quality of the prompt overall. If you see that the guidance is not being properly applied or follow, the first step to troubleshooting should always be to optimize or reword your guidance prompt. And now I'd like to show you how guidance works within the framework of our trademark AI engine. So this is the AI engine that you see on the screen here. You've probably seen it from previous webinars or our website. This is the step by step process in which our AI engine ingests an inbound query, so a question from a customer, and then generates a suitable response all within a matter of couple of seconds. We're not gonna go through each step in detail, but I just wanted to show you where guidance is actually applied in this process, which is gonna be around here. So most guidance is actually applied once the AI engine begins the response generation phase, where it combines all refined queries, retrieved content, and reranked information in order to produce customer facing responses. It's at this point that any relevant enabled guidance used to control the generated answer will be applied as the AI agent finally determines if it needs to generate an answer, ask for clarification, or take an action. And the reason I specified most guidance is because we've expanded the classification of guidance recently so that it's actually applied at different phases throughout the process based on the guidance type. Although we have multiple buckets outlined within the platform, we can simplify the guidance categories down to three main classifications. So first, we have guidance to detect, escalation situations. Then we have guidance to draw from specific knowledge content. And then finally, general guidance, which just helps guide Finn's overall conversation style. So for an even clearer view of how these classifications interact with each other, This is how the different types of guidance are applied throughout the entire AI engine process. So first and foremost, escalation guidance is taken into consideration. Therefore, determining whether the AI agent should handle this issue at all or to hand off to a human agent. Next, we have content guidance, which can help speed up the re ranking phase when determining the most relevant content to the initial query. In fact, it's gonna limit the amount of content that the AI engine has to retrieve from, which is going to improve both latency and then accuracy as well. Finally, any communication or behavior based guidance will be applied to the generator phase when the AI engine begins to produce a response from the retrieve content information. And then with the final response, you'll be able to see all the guidance that was successfully applied by toggling the conversation events, which allows you to see whether or not it was applied appropriately. Another concept that I'd like to clarify for the group is just how guidance should be treated. So overall, guidance will influence Finn's demeanor and decision making based on the instructions we provided. We can influence Finn's tone of voice and answer length, of course. This was one of the first versions of guidance that we rolled out, early last year. There's also situation and audience based guidance. So how Finn should speak in specific situations or towards specific customer segments. There's content and source guidance as we've discussed, so which sources Finn should reference for specific topics or situations. And then finally, situations where Finn should not be involved in at all and escalate to a teammate. But I think it might be more helpful to provide examples of how guidance should not be used for. So, for any operational actions that are typically accomplished through workflows, you need to continue relying on those in order to execute. This is not what guidance can be used for. And on a point of clarification, escalation guidance will route to a default team, but cannot route to a specific team just from guidance. This will be determined by how your workflow is structured, which I'll cover later on in today's session. Negative content guidance is also unreliable. So for instance, telling Fin not to use a specific source for certain scenarios, it's a lot more reliable if you just remove the content from Fin's usage altogether. And as well, you can't influence Fin on what it doesn't know yet. So for example, if you just tell Fin to escalate if they don't know the answer, the escalation guidance is applied at the beginning of the AI engine process as we saw. So the instructions can become counterintuitive because Fin doesn't know if it knows the answer or not in this specific scenario. So tying it all together, guidance only impacts Fin's ability to escalate and its conversational behavior. If you need your AI agent to collect, read, or update data for personalized queries, then you'll need to rely on data connectors and tests. It's best used for informational queries that are single term. Once multiple step turns, or or multiple steps are introduced, then guidance just won't be of much use anymore. Alright. So now let's get into some more practical advice provided directly from our AI team. We'll start off with some classic troubleshooting advice. So very often, we're asked by customers why guidance isn't working or being applied. Again, this all goes back to AI prompt quality and the relevancy of the inbound query. So firstly, the let's go through each step. The first one being that guidance might not be properly enabled. So, this one is just like for dummies. Of course, like you have to both save and enable the guidance just to make sure that works. If you've already gone through that, next step might be to, consider that maybe your guidance might be too specific to a particular scenario and that's the reason why it hasn't been triggered yet. Fin needs to recognize when to apply the guidance based on context clues within the conversation in order to execute. Thirdly, the customer's query might not contain the keywords that trigger your guidance. So make sure that your guidance is thorough enough that it captures a wide range of possible questions and conversations. Next, your guidance might be too broad or mixing multiple or conflicting instructions. Very often I've seen from users, that they try to do, all encompassing piece of guidance that nails everything. But we wanna do is split your guidance into specific objectives. So each guidance piece should focus on almost like a single mission. Don't try to clump too many into one piece of guidance alone. Lastly, the customer may not match the audience rules that you've applied to your guidance as well. So reexamine your audience rules and making sure that, when you are conducting your testing, that you test with those specific personas to see if it matches or not. When you're testing guidance, try doing it in the preview panel with sample questions that you expect to trigger the guidance or even using real questions that you receive from your customers within your inbox. You should always select the preview user drop down if you need to test whether guidance is being triggered correctly for a specific audience or user. And then when trying to reword or optimize your guidance instructions, make sure that you've selected the optimize button just to see if the structure or placement of your guidance could be improved. The optimize button is also going to help spot any contradictions, within your guidance pieces. As you see on the screen here, I've asked it to review a piece of guidance of mine, and it it identified immediately a potential contradiction with another piece. And then finally, sometimes splitting guidance into separate sections can help Finge trigger them in the right context. Again, instead of bundling too many of them together. Alright. Now let's dive into some escalation guidance and some of its core principles. Essentially, we can split escalation guidance into two main concepts. So first, we have direct escalation, which is escalation in its most traditional sense. Therefore, if Tim detects a suitable opportunity to escalate, it will do so immediately without any turns in between. Then we have the offer to escalate, which is rather new with escalation guidance. So we can now coach Tim to provide verbal offers to escalate within a message directed at the user. And only once it's confirmed will Tim then proceed with the direct escalation. By its agentic design, Fin, by default, is gonna try and analyze certain elements within a conversation, to inform its decision to employ escalation guidance or not. And I'm gonna walk you through some of these elements now. So if someone straight up asks or requests to talk to a human agent, Fin's gonna escalate immediately. This, however, can be modified with guidance. So for example, Fin can be asked to offer escalation instead of directly escalating, or Tim can even be asked to refuse escalation too, depending on, the situation. For keywords, at the moment, if user types a word like human or agent, Tim already gonna offer to escalate, or it's gonna clarify if they'd like to talk to a human agent or not. You can also play around with this using guidance, asking Tim to adopt different behaviors for different keywords. So some examples would be, like, ask Finn to escalate immediately if the account owner's name is mentioned or let's say, if the the customer mentions a competitor. For things like anger and frustration, Finn's also gonna offer to escalate. And with guidance, you can modify these settings too. So for example, you can ask Finn to look for the slightest signs of frustration, which would in return make Finn a lot more sensitive. It's also good at identifying loops as well. So if the same answers are created three times in a row, it is recognized as a loop for Finn, which in return, it's gonna offer to escalate. You can modify this with guidance by having Finn offer to escalate after two times or you can even ask Finn, for example, to vary the wording of its responses each time to avoid super repetitive answers. And then finally, we have negative feedback. So Finn's offer, Finn's going to offer to escalate. If a user gives negative feedback without additional information twice in a one conversation. So negative feedback includes all comments or replies that can be considered negative in connotation or sentiment. So things like passive aggressive nos, for example. These negative feedbacks won't even need to be consecutive. They just need to occur twice, at any point in the conversation. But once Finn detects it twice, it's going to offer to escalate. So, using guidance, you can also modify this setting as well. So you can offer, to escalate after the first piece of negative feedback, that fin detects. As I mentioned earlier, guidance itself won't be able to route to a specific team. However, used in conjunction with branching logic, you can determine different escalation pathways within a workflow by selecting a specific piece of escalation guidance from the drop down menu. They should now be appearing as attributes you can select when building out branched pathways. And this can be ideal for many different use cases. So for example, if it detects specific sensitive topics, can now route the conversation to specialized teams that may have expertise in their respective topics. And now we're just gonna go through some general best practices as you write out your instructions. So always refer to Fin as you as if you were training a teammate or a new support agent. What we found effective as well is to set expectations within by giving it a role or a job at the beginning of your guidance. So for example, you can tell Fin something along the lines of, like, you are a coach and your job is to improve the customer's performance. Kinda just sets the tone for the rest of its response. Next, you wanna make sure that you're giving complete instructions when writing guidance as well and keeping it as concise as possible. So although you have, 2,500 characters available per piece of guidance, the more simple it is for a human to understand, the more it will be for your AI agent as well. If you wanna reinforce different scenarios, include both positive and negative examples for the AI agent to reference, but avoid using too many connector words like and and or. Each piece of guidance should also operate individually to serve a unique purpose, which is why we recommend dividing up your guidance by subject or task. Again, treat each piece of guidance as if they had a specialized job to carry out that doesn't contradict other guidance pieces. Guidance structure actually matters a lot. It's key in terms of having your AI agent understand it and then follow it when appropriate as well. So to start, use, I guess, like a structure or sequence that we often recommend. So first start off with a sentence or two to set the context like giving it a job description or setting the scene of the parameters in which it's going to operate. And then describe the specialized objective or or the the task that it's seeking to accomplish. So remember, keep this concise as well. Follow that up with do's and don'ts. So provide both positive and negative examples of scenarios where this piece of guidance will apply. And then finally, you can provide direct quotes or message formatting examples that the AI agent should use in its reply. Again, the realm of possibilities is endless here, so there's lots of room to innovate in this field still. We'll be covering different use cases, for guidance in the demos that are gonna follow later on in today's session. You should also understand technical terms that Fin already knows, to execute when you are running out some of your guidance prompts. So we've included some of these terms below, and a lot of these we've covered already in today's session. So, escalation language, for example, escalate and offer to escalate. Tim gonna understand what that actually entails, direct answers for any, like, informational, queries that it's gonna provide, and then positive and negative feedback, of course. So Tim really good at detecting sentiment from customer messages. As well, be direct and precise in your your language when you're writing your prompts. You don't need to add a lot of filler words or be overly polite with the AI agent as well. Trust me. It can take it. Just be direct, and and stern when you're addressing your AI agent. As well, when you are offering alternatives, for example, don't leave it up to the LLM. So when you're instructing it don't do something, you want to always provide an alternative of what it should do instead. Otherwise, the LLM is going to have to figure out on its own what to do in that negative scenario. And this one's pretty straightforward as well. There's no way Tim can prioritize one piece of guidance over another for the same audience, so try and avoid contradicting pieces of guidance. As we've demonstrated earlier as well, the optimize button is actually gonna help you spot contradictions within your guidance while also providing recommendations for improvement as well. So make sure you take advantage of that button for your instructions. And if you need to emphasize certain elements to really drive the message for Tim, you can use all caps, to let the AI agent really know the significance of a particular element of your prompt. And finally, you can use guidance to also dictate how Fin should respond in certain scenarios word for word, using quotation marks. So this might give you better control over specific events during a conversation, like either a greeting or a handoff message to the customer. Otherwise, the response will be generated by the designated LLM if you leave it to Fint. You can also pair dictation with, handover guidance, for example, if you want Fint to perhaps ask clarifying questions before handing a cost a conversation off to a teammate. And this is gonna make for a lot smoother of a handover while also collecting more context for the teammate assigned. Alright. So now for the demo, I wanna walk you through a quick demo in my test workspace real quick just to demonstrate some, guidance structure that I've used to to write a specific piece of guidance. Then I'm gonna hand things off to Nathan Suds, who's a top community expert here at Intercom. He's gonna walk us through some really cool, use cases of guidance that he's used recently. So without further ado, let's hop into the test workspace. Hey, everyone. I'm on this side of the screen now, and we're in my test workspace. I just wanna show you a couple pieces of guidance that I wrote and how they're gonna structure the response that our AI agents can provide. So, the one I wanna focus on today is about chess openings. So I've been playing a lot of chess recently, and a new opening that I'm trying to learn is the Sicilian defense. So, this is an example of a piece of guidance that I wrote, to have our AI agent provide a lot more, advice on chess openings, in case a customer asks about it. So we wanted to detect a certain topic, which is chess openings, and, we're gonna follow along the format that we covered in today's session. So we give, the agent a brief description, followed by, basically formatting the response in a specific way. So we're bolding specific text. We're creating, like, little subheadings. It's almost sequential. So we're we're telling it what to write, for example. So I want a brief description of the opening, the first four moves, for example, which we outlined here, and then a list of pros and cons with the specific emojis. And then finally, I want a list of games where, grand masters have played this opening with, links to these specific games hyperlinked. I also have another piece of guidance activated where, when it's sharing URLs, that is gonna be hyperlinked within the text and not providing the URL in full. Finally, what we'll see in my response is that it's always gonna start off its responses with the knock knock joke. So, for example, let's ask it about the Sicilian defense. It's an opening move for black that I've been trying to learn recently. And let's see. There's our knock knock joke. And as you can see, our AI agent has done really well at following the formatting that we provided in our chest opening structure piece of guidance as well. So everything's correctly hyperlinked as well, to our specific games or the specific, players that have used it. So it could still use some fine tuning. I did ask it to link, like, specific games, but it's just listing, I guess, grand masters that have used this opening. And if we go to our event log, for example, it is going to show us all of the guidance that was successfully applied. So we have our knock knock joke one. We have the one about chess openings, and then we have the one about URL hyperlinked as well. So this opens up, I hope, a realm of possibilities in terms of how you want your AI agents responses structured. Use the communication style section for guidance on this one. And now I'm gonna hand things off to Nathan Suds, again, who's a top Intercom community expert with us, and he's gonna showcase a lot of, like, cool use cases for guidance that he's used as well. So I'm gonna toss things over to him now. Everyone, it's Nathan. You might recognize me as the top Intercom expert at the Intercom community. I'm over here today showing you some creative stuff with Fin, Fin Guidance specifically, and with this creative company, Jukebox. We did some work with this client, to bring their creative ideas with Finn to the creative, work of, like, print work. So we've this this company ships die cut stickers and business cards and creative work all across The US and Canada. If you need business cards or stickers or anything like that, they've got everything you need there. And, of course, with that comes a lot of support, and automation systems that are, can really use optimization with intercom. So, basically, what we did, that I thought was really interesting and was talking to Tim about it. And so wanted to show this creative use case where people wanted to get an order history. So they wanna reach out and say, hey. Can I see my recent orders? Or I wanna reorder x y z, item that they had. And the challenge we're having is we're using a fin task or or custom actions, for example, and we wanted to get an order history that comes back that looks kinda clean and simple, something you might see from, like, Amazon or, Shopify or something like that. But they're using a custom back end, a custom shopping cart of their own, and so it didn't have these kinds of things built in. And so what we did was we connect the API with Fin using, Fin tasks. So the, actually the data connectors and custom actions. And then we got the results back, but the results were just this kind of plain text based response, no real flare to it. And I tried as much as I could in the, actions to get it to work. It just didn't seem to, like, give me the flexibility that I wanted and the customization. And then all of a sudden, I had this idea of what about guidance? I wonder if I could use guidance to, like, design the response from Fintech or from the custom actions. And so I had this moment, went into, custom I went into the guidance and I started working on it. And took a little while, not too long, but took a little while to kind of finesse it. But I got it down to a really nice prompt where we basically said one key thing here was we said we always wanna keep include the links. And I'm kinda design focused, so I wanted to make sure that it didn't just drop a link in there. I wanted to say view order and then have a little emoji next to it. And I also wanted to show the status and the, you know, awaiting files or whatever status it was. And so we kinda worked around a little bit with this design at the way we wanted to and gave Tim a few examples of, like, projects with multiple products or something with quantities in it and just kinda represent most of the kind of scenarios that we would have. And then I, worked with chat GBT and clog to come up with this framework where it's like, okay. Here's the different status emojis. So these represent the different statuses that a project could be in. And, these are really important for the project overall, but, you know, at a glance, really quick easy view for the customer was the real focus here. The real powerful back end with Tim is in the custom actions where you can actually we can actually tell Tim if it's in production, you actually can't cancel this order. If it's if it's preproduction or if it's in files received or something else, then you can go ahead and cancel it. If it's already shipped, you can't update the add address, that kind of stuff. So we work with Tim on that on the data connector side. But this was more important about the order history. So all these statuses are really important because the customer needs to see, like, if it's in production, sorry, we can't cancel the order or change the order now. And, so we just have this prompt that basically says, follow this. Here's the desired output. Here's an example. And then here's the example of all the different statuses that also, are needed. And then we just said, if you include any instructions about how to view your view your orders in the customer account center, do that below the order history. So the order history is the focus because what we're finding was Tim was like explaining it first and then giving the order history and we didn't really like that as much because people were just asking for the order history. There's some scenarios where it might do the other way around sometimes if that's what they're asking for, and then it shows in the order history as well. But in those cases where we, you know, the real focus was the order history. We wanna make sure that was first. And then we said, here's a link to an individual order in the history. So I wanted to show an example of what an individual order link looks like, and then I showed it an example like order number, replace the order number with the actual order number. So if Tim gets an order number at some point, it could look up an order number and put it in the URL and get it linked to the customer. And so this is, you know, kind of a creative prompt in my opinion to get dashboard links and order history to the customer. And, I just think, like, when you think outside the box, you start to realize you've got lots of things you can do here. And so anything you wanna shape or transform or style in in your responses from Fin, you know, think about it this way. You could really do it. And, let me just jump over to the tasks area, the action builder so you can see what I'm talking about here. So we have a number of, FIN actions in the data connectors. So cancel order by order number. When a user asked to cancel, check first to see if it's eligible. And then get order tracking shipping information that's related to the order. And then there's the rush order. So we actually have a really cool functionality here as well, where if somebody asks about an order, because the team is perfectly fine with rushing an order, if someone reaches out and asks for something, in a certain way, we can actually put a rush on the order automatically. So Fin can recognize that, put the action, put the rush on it for us. So that really helps us to speed up the process for customers when it's when it's important. And then we have get, customers recent order history, and this is where that whole, magic happens. And then we also have get a reorder link, which is just one reorder link for a specific order, and then we have, you know, get the order by the order number. So if they throw the order number in there, Finn, Finn can look it up and and find out about that. So if we go here into the get custom order history or our customers recent order history, you can see that I stylized it here and I had all that same information here. But when I did it here, it just didn't it just didn't do it. It didn't apply it. We would get it back in more of, like, a format kind of like, like this kind of thing. And we tried tweaking it, tried removing stuff, but this was, like, the end result that worked worked for us. I put it here and I also put it in the other space, but but it wasn't until I put it in guidance that it actually really, really shaped it the right way every time. Yeah. And so that's a nice nice example of how you can put the custom actions to work with the guidance to make things happen. And let's just go into the inbox and show you one of these in action. So this person, Anna here, reached out and said, look at this. Even said my order numbers are and listed all the order numbers. And this is actually if you notice, this is by email. So we have Fin over email running. And this person, I wanna see you. It could be at, like, late at night. This is a 04:44 forty PM. So this might have actually been, like, right before closing. They're checking on this. And the team may have not gotten to it in time. But in this case, they didn't even need to because Finn picked it up and said, hi, Anna. Here's your order, you know, here's your information. So they showed all the orders and the status of each one of those orders, and then he said, all five orders are currently in shipping status with an estimated delivery date of October 2. And I don't have full information about the specific tracking numbers yet or or order carrier details. These orders are part of a group shipment being sent to your Toronto address and then it sells it that's composed by AI. And so this person got the information they needed at 04:45PM right before closing without the team even, you know, getting involved. But they also didn't have to go to their my account page, log in, look at that information. They could just send the order numbers and get a response. And this, like, again, this is by email, so they didn't even visit the website. They just sent the information over. And so there this is, you can see here, if you have, the show conversation events on. So if you have it turned off, it's kind of, no information there. But if you turn on the show conversation events, you can see how Fin went and checked the order tracking information, you know, and got all the information for each of those orders and then pulled that back and then, use my guidance here on order history and dashboard links to create the the view, for the customer. And so this is how Finn in action for any of your ideas can really, like, shape the customer experience better and, get creative with it, you know. I'm look I'm looking forward to hearing how you use it creatively after this, this webinar. Thanks a lot. Good luck. Thanks thanks and forth, the the cool video that you sent over again. I really, really appreciate it, and hopefully, we can see more as we continue to innovate with guidance. One last thing before, we move on to q and a. I do want to plug our upcoming flagship event. So we'd love for you all to join us at Pioneer twenty twenty five on October 9 happening in New York City, both in person and virtual. If you don't know what Pioneer is, it's Intercom's biggest event of the year, and you'll be the first to see the unveiling of some exciting new product updates from our leadership team, learn how real companies are using AI to change the way they support their customers, And then you'll also get the chance to connect the network with other CX leaders who are shaping the future of customer service. So, these product updates are gonna be brand new. So, you and I will both be learning about, potentially, like, some some new product announcements at the same time. So I personally am super excited for this. I've included the link in the docs tab, to sign up, so make sure you register there. Otherwise, I will move on to q and a now and invite my colleagues to hop on stage and, answer any remaining questions. So make sure you upvote those questions in the q and a tab. Thank you. Okay. Just checking on, sir. I'm I'm not muted. K. Yeah. I'm said colleague backstage. I do wanna share some questions that, that were asked that I think were really good. A common one that a lot of people were asking about today was about, a lot about, like, different brands, for example. So how is Finn going to adhere to, brands? So multi branded Finn is currently in closed beta right now. But if you reach out to your relationship manager, they can opt you into that. I'll share the question. Yeah. So you can get opted into that, and it's all based around audience rules. So, when you're building out your audiences, for example, you can apply different attributes to that. That's how Tim is gonna be detecting, like, which brand, it is it's gonna adhere to, and the same rules are gonna apply to your guidance as well. So once it detects that, people are interacting with, like, I don't know, a certain brand of of Tim, for example, it's going to adhere to all of the the relevant guidance pieces there. There was another good question that I saw about, someone asked about, like, an example of when it'd be useful to, like, offer it to escalate, for example, and this is the question here. This, again, is for how mature your AI agent is, and it's for your own use. Okay? So when you give an escalation guidance, it's going to be maybe a bit trigger happy. And so maybe it might detect, like, any, I don't know, mention of chess, for example, you should escalate. So it's gonna be, like, super, super trigger happy. But if you just, offer instead, what that's gonna do is it's gonna free up more of your team's time, and hopefully lead to more deflection and then resolution, of course. So that's when you should instruct Fin to to offer escalation. If you prefer to lean more heavily on your human, support, then, you know, you can just disregard the offers and just, like, escalate immediately. So that's that. Let's see what other questions, that were asked recently. So this one might be cool. So you can use data connectors to get really customizable with customers. What are the limitations on this? Great question. So I think I had a slide today about, like, different use cases for guidance, versus, data connectors and tasks. The only limitation I can think of with data connectors is, you can simply, like, read and and write data. Right? So you'll leave both a poll, let's say, someone's account status, someone's account number, for example. But it's only, like, pulling information like that. So it's not gonna be useful for, like, multistep questions, for example. That's where we're gonna need to lean more on tasks, data connectors. So, like, there could be, like, increased latency. I guess, like, that's another limitation I could think of. But each one kinda has, like, a different use case, and, I'll try to find that help center article, to share with you, Greg. Otherwise, like, I don't know. Like, it's it's hard to find limitations data connectors. I think they're great, especially when you look at your conversation taxonomy from the the insights tab. You're gonna wanna analyze, for example, how many of your questions are coming in, require a bit of personalization. And once you set those up, it's gonna be easy peasy in terms of, addressing, like, very, very personalized questions from customers. So, yeah, let me find that, help center article for you. I'm gonna ask my colleagues backstage. There's a cool article that defines, like, okay, one d's data connectors versus guidance versus tasks. Yep. Also, again, like, tomorrow's oh, sorry. Next week's event at Pioneer, they are announcing a lot of cool new features that I don't even know what's being announced or released yet. So let's wait and see. Like, a lot of, we got a lot of, like, good suggestions on, like, product feedback, and and different use cases for these features I went through today. I think they're gonna be making really, really big announcements, next week. So I'm not even trying to to to, like, hype it up too much, but, like, I myself am excited to see what's gonna be announced because they haven't even told us yet. Thank you. Rigby. Yeah. Thank you. That's the article I was thinking of. I think another good point that I want to bring up was something that that Nathan mentioned was, like, relying on LLMs like GBT or Cloud to help you write some of these prompts. Someone asked what the optimized button actually looks at. It's hard to say, but it does rely on an LLM for, like, best practices. So it honestly, what it's doing is it's calculating, like, how can I word this best for me, the the the LLM, to to understand? So it's making it better formatted for AI to understand, and recognize and execute. And on top of that, it analyzes for contradictions. Some of the best practice that we shared today, are more like human focused. Right? So, when we talk about things like, empathy or even phrasing, you're gonna still want to rely on humans to write, those specific instructions. But things like formatting, and and, like, contradictions, for example, it's gonna it's gonna it's it's gonna help you out with, like, the technical aspects of guidance, basically. Can we input all of y'all's instructions for guidance into GPT and then ask it to write us a response? This could be helpful. The optimized doesn't look at everything. Yeah. Again, like, for best practices for, like, prompt writing, I think what I mentioned earlier in today's session, so I give it always give it a a a task at the beginning. So let it know that, like, you are writing, I don't know, prompts for, an AI agent guidance, and this is what you're looking to achieve. Once it does that, it's gonna better understand, for example, like, what the the main objective is and then how you want it, like, formatted, for example. So you could try and let us know how it goes. I've never thought of just adding all of our content today into to, like, an LLM and having it spit out results. But it depends on what kind of guidance you're writing as well, obviously. So, you could try it. You know? Let us know. Share in the community. I linked in one of the answers today, but I guess I'll plug it one more time, and maybe I'll add it in the docs tab next time. But, the guide, the the Intercom community, this is folks, this is where you can submit, like, product feedback, product ideas, for example, and also just network with our community of, Intercom users and fin users. A lot of people are gonna share best practices, so you can, like, network with them. Yeah. Of course, Nathan's gonna plug it. Like, he's our community guy. But network with other fin users, but also network with other, like, industry peers, because they might have a lot of, like, similar use cases, and then you can share best practices with each other there, if not through the the webinars. Yeah. Yeah. We've done, like, a a lot of, of work on the community recently, and so it was it was dead for a while, but, like, it's really, really bounced back. So I guess we can thank Nathan for that. Alright. We have two more minutes, folks. So if you have any other questions, we'll stick around. But, yeah, I would really, really recommend the the formatting tip that we we went through in the in the demos today. I just came up with that demo, like, last night, and I was pleasantly surprised at how well it worked. The hyperlinking one could use a bit of tweaking, but that's one that I stole that actually directly from, Intercom's, our own support team. That's a really popular one I've actually seen in a lot of workspaces. So, again, simple formatting like that, like, asking Fin to, let's say, like, always hyperlink URLs, to bold specific words or italicize. Fin's also really good at doing that. And, actually, I there was another good question that I wanna share where someone asked about, like, the the depreciation of, custom answers. So in in lieu of custom answers, what we recommend, is you making really, really good use of those quotation marks if you're trying to have really exact phrasing. Oh, here it is. So for, like, really exact phrasing or specific actions even. So for actions, of course, like, you'll need to rely on the the data connectors, anything that you set up there. But, using quotations will really, really help with, being very deterministic and and exerting a lot more control over, VIN's response generation. It it could be the reliability. It's it's still a bit wonky, but I don't know. If you add clear clear emphasis, to use specific phrasing and, like, specific scenarios, Finn will be following those instructions to a team. So yeah. Alright. I guess we're gonna wrap things up for, today. Hope we're not missing oh, sorry. You know what? Let's do one last question. Is it advised to use markdown formatting to write the prompts and guidance? I've noticed it is more effective in some agents. I'm not familiar with the term markdown. I'm sorry. If you think it helps, then that that's great. Maybe I'll follow-up with you on this. I'm not sure what what markdown formatting refers to. I'm so sorry. But, yeah, let's wrap it up for today. Merrick, I'll I'll, follow-up with you after today's session. I will be sending out the deck to all the live attendees, and the recording will be shared with anyone that's registered. So thank you all so much again for, joining us today. Really appreciate you coming, and, we'll see you at Pioneer next week, October 9. Bye.